| Summary: | 碩士 === 義守大學 === 電子工程學系碩士班 === 96 === H.264 is newest video coding standard and it can achieve higher compression efficiency compared to previous standards (such as MPEG-4 and H263) in the same video quality. To achieve the target of high video quality and compression ratio, H.264 employs some new techniques. Two important approaches are the techniques of variable block sizes motion estimation (ME) and sub-pixels motion vector accuracy. H.264 uses 7 variable block sizes ranging from 16?16 to 4?4 in inter frame coding. The integer ME module with 7 modes and sub-pixel ME (SPME) module need very high computational complexity. In other words, the ME module is the most time-consuming component in H.264 encoder. Therefore, it is difficult to real-time applications for H.264.
To reduce the complexity of ME module in H.264, we propose an efficient SPME algorithm based on exploiting the high correlation of the neighbor blocks to find the prediction motion vector (PMV). First, we use the PMV to predict the corresponding motion vectors (MV) of 7 modes. If a block it is likely to have a small cost at MV(0,0). Therefore, we define the thresholds (THZs) of zero-motion blocks for seven block sizes to determine if the ME will be early terminated. The THZs for different block sizes is selected according to experiments on many sequences. In mode 1, since one macroblock (MB) contains only one 16×16 block, if the zero- motion cost is smaller than the mode threshold (THS), the 16×16 mode is the best mode of this MB and the ME is stopped. In addition, in order to further speed up the encoding time and improve video quality of H.264, we propose an efficient SPME algorithm based on combining the methods of fast ME and fast hierarchical SPME.
Experimental results show that the proposed efficient SPME method can achieve an average speedup factor of 52 and 13 when compared the fast full search (FFS) embedded in JM 12.4 [19] and the uneven multi-hexagon search (UMHexagonS) algorithm [13], respectively, with insignificant loss of image quality.
|
|---|
